Can we manage first impressions in cooperation problems?

Transcription

Can we manage first impressions in cooperation problems?
Preprints of the
Max Planck Institute for
Research on Collective Goods
Bonn 2011/5
Can we manage first
impressions in cooperation
problems?
An experimental study on
“Broken (and Fixed) Windows”
Christoph Engel
Sebastian Kube
Michael Kurschilgen
MAX PLANCK SOCIETY
Preprints of the
Max Planck Institute
for Research on Collective Goods
Bonn 2011/5
Can we manage first impressions in
cooperation problems?
An experimental study on “Broken (and Fixed) Windows”
Christoph Engel / Sebastian Kube / Michael Kurschilgen
March 2011
Max Planck Institute for Research on Collective Goods, Kurt-Schumacher-Str. 10, D-53113 Bonn
http://www.coll.mpg.de
Can we manage first impressions in cooperation problems?*
An experimental study on “Broken (and Fixed) Windows”
Christoph Engel
Sebastian Kube
Michael Kurschilgen
Abstract
Cooperation problems are at the heart of many everyday situations. In this paper, we propose a
very simple and light-handed mechanism to sustain cooperation and test its performance in a rich
laboratory environment. The mechanism moderates cooperation by controlling experiences,
more specifically, it "manipulates" subjects’ initial beliefs by providing them with selective information about (un)cooperative behavior in other, unrelated, groups. We observe that contributions are considerably sensitive to such selective information. First impressions participants happen to make predict subsequent behavior. Our results, however, suggest an asymmetry in the
strength of the reaction – which might pose a limit on the effectiveness of the mechanism in natural settings.
JEL: C91, D03, D83, H41, K14, K42
Keywords: Public Good, Behavioral Uncertainty, Conditional Cooperation, Information, First
Impressions, Broken Windows
*
Helpful comments by Aniol Llorente-Saguer and Matthias Lang and by the audiences at the Annual Conference of the Royal Economic Society, the Association for Public Economic Theory, the International Meeting
on Experimental & Behavioral Economics in Bilbao, the University of Bonn and the Max-Planck Institute for
Research on Collective Goods are gratefully acknowledged. Lauri Sääksvuori was kind enough to run for us
the belief treatment in Jena. Darius Wieczorek programmed the experiment in zTree.
1
1
Introduction
Cooperation problems are at the heart of many everyday situations. For example, when it comes
to protecting the environment, defending one’s country, generating new knowledge, joining a
political party, extending the infrastructure, or exploiting the opposite market side, agents face a
social dilemma. Jointly they are best off if everyone contributes her fair share. But individually,
free-riding on others’ efforts yields the highest payoff. A large number of theoretical and empirical papers have explored how people should be expected to and how they actually do behave in
such situations. Many empirical contributions make use of laboratory experiments where subjects participate in a prisoners’ dilemma or in a public-good game. Absent institutional interventions like punishment (Fehr and Gächter 2000), contributions are heterogeneous in the beginning, but average cooperation quickly declines and most participants free-ride in the end (e.g.
Andreoni 1988). These results have been stress-tested extensively, e.g., with respect to anonymity (Andreoni and Petrie 2004), culture (Herrmann, Thöni et al. 2008), group size (Isaac, Walker
et al. 1994), efficiency (Glöckner, Irlenbusch et al. 2010), or framing (Goerg and Walkowitz
2010). Interventions that mitigate the dilemma are not easily designed. Effective interventions
tend to be heavy-handed, often altering the incentive structure of the game such that free-riding
is not in an individual’s self-interest any more (e.g. Ostrom, Walker et al. 1992; Falkinger, Fehr
et al. 2000; Fehr and Gächter 2002; Gürerk, Irlenbusch et al. 2006; Glöckner, Irlenbusch et al.
2010). In this paper, we propose a very simple and light-handed mechanism and test its effectiveness by using laboratory experiments.
Our mechanism is based on an observation that has previously been made in public-goods experiments, namely that (initial) group composition and initial cooperation rates significantly affect
the future development of cooperation rates in a particular group (Burlando and Guala 2005;
Gächter and Thöni 2007; Beckenkamp, Engel et al. 2009). A likely reason for this behavioral
pattern is provided by the influential concept of reciprocity and conditional cooperation (Fischbacher, Gächter et al. 2001; Dufwenberg and Kirchsteiger 2004; Falk and Fischbacher 2006;
Fischbacher and Gächter 2010, among many others). If persons are sensitive to the behavior of
other group members, those groups who cooperate little in the beginning will become even less
cooperative over time, while groups with substantial cooperation in the beginning are able to
sustain cooperation over time. This suggests that outsiders might moderate cooperation by controlling experiences. Our mechanism is even less invasive in that it is confined to the first impressions subjects happen to make. More precisely, the idea is to “manipulate” initial beliefs by
providing participants with selective information about (un)cooperative behavior in other, unrelated, groups. The information affects initial beliefs about others’ behavior, which in turn changes how I myself behave initially. A virtuous (or vicious, depending on the kind of information
that is provided) cycle starts: via beliefs, our manipulation determines cooperativeness in the
beginning, which in turn determines cooperativeness later on.
Our intervention idea is in the spirit of James Q. Wilson’s Broken Windows Theory (Wilson and
Kelling 1982) which is based on the work of American sociologist Philip Zimbardo. In 1969,
Zimbardo abandoned two identical cars in two different locations: the Bronx, NYC and Palo
2
Alto, California. “The license plates of both cars were removed and the hoods opened to provide
the necessary releaser signals” (Zimbardo 1969). In the Bronx, the abandoned car was stripped
and demolished after only 26 hours, the result of 23 separate incidents of vandalism. In contrast,
the car in Palo Alto still sat unmolested after the course of an entire week. Zimbardo then decided to provide an example of vandalism to the affluent and seemingly non-violent neighborhood
of Palo Alto. So he and two graduate students of his took a sledgehammer and started bashing
the car. After they had taken the first blow, observers shouted encouragement and finally joined
in the vandalism, until the car was completely wrecked (Zimbardo and Ebbesen 1969).
In order to clearly identify the short- and long-run effects of our mechanism we run laboratory
experiments, which have the benefit of providing a sufficient degree of control over the environment. The environment that we use has been prominent lately to study social dilemmas and
cooperation problems. It is a complex public-good game (more precisely, a voluntary contribution mechanism) with decentralized sanctions (Fehr and Gächter 2000) and counter-punishment
opportunities (Nikiforakis 2008). We have chosen this rich environment, since it mimics several
potentially important features available in natural environments. Moreover, it is sufficiently
complex, so that conflicting behavioral norms might emerge – which allows us to test the effect
of both favorable and unfavorable first impressions.
In the baseline, subjects play the game for 10 periods without any exogenous interventions. In
two treatments (FAV and UNFAV), prior to making their first contribution decision in the game,
we provide subjects with selected data from the baseline. Specifically, we show them graphs of
the development of mean contributions over time from selected cooperative (resp. uncooperative) groups. The information that we give is unfavorable for cooperation in treatment UNFAV –
similar to Zimbardo, we “attack the car with a hammer” and explore whether “the car will then
go to ruins”, i.e., whether cooperation breaks down in the public-good experiment. Contrarily, in
treatment FAV we reverse the manipulation by exposing subjects to favorable initial impressions
– i.e., similar to Wilson, metaphorically speaking we repair the windows in a bad neighborhood,
and investigate whether newly arriving inhabitants behave more cooperatively.
We find that cooperation rates strongly react to the treatment manipulation.1 Unfavorable information substantially reduces initial contributions to the public good. Moreover, the effect is longlasting and translates into future development of groups’ cooperation behavior. Contributions in
treatment UNFAV do not recover during the experiment. More importantly even for policy making, the favorable selective information in treatment FAV prevents cooperation rates from decaying over time. Taken together, our results suggest that subjects’ behavior is indeed conditional on
initial expectations of how others are likely to behave, and that these expectations are open to
purposeful intervention.
1
As we demonstrate in a second series of experiments (reported below in Section 4), the effect of our manipulation is mediated via subjects’ initial beliefs. Beliefs about others’ initial contribution decisions are much
more pessimistic when unfavorable information is provided.
3
The next section embeds our study into the existing literature. Subsequently, we explain the design of the experiment. Section 4 presents and discusses the experimental results of our main
study, and features an additional experiment that explores whether the effect of our manipulation
is mediated by subjects’ initial expectations. Section 5 concludes the paper.
2
Related Literature
Our study closely relates to the experimental literature on reciprocity and conditional cooperation. This literature suggests that a substantial fraction of subjects usually consists of conditional
cooperators, i.e., people who are willing to cooperate, provided a sufficiently large fraction of the
population does the same. (Fischbacher, Gächter et al. 2001; Mengel 2007; Fischbacher and
Gächter 2010). For conditional cooperators, information about contribution patterns in other
groups has a first- and a second-order effect. Given the uncertainty about the composition of the
group of which they happen to be a member, information about other groups helps them form
beliefs. Moreover, each participant knows that each other participant has received the same information. This makes it possible to also form a second-order belief, based on knowing that the
remaining group members have received the same information. The finding that, in our experiment, cooperation rates are sensitive to the initial information-manipulation is fully reconcilable
with the idea that a non-negligible fraction of the population acts as if they were conditional cooperators. This yields further support for the existence and importance of conditional cooperation.
The research question covered here is also closely related to the literature on institutional design,
specifically the branch that explores interventions aiming at raising contributions to public
goods. Usually, the proposed mechanisms alter the incentive structure such that free-riding becomes less attractive, or even a dominated strategy for self-centered money maximizers. For example, in (Falkinger, Fehr et al. 2000) the payoff structure is changed such that “each individual
gets a reward or has to pay a penalty depending on the deviation of its contribution from the
mean contribution.” In other studies, group composition is changed such that the payoff structure
is changed implicitly. For example, in (Gunnthorsdotir, Houser et al. 2007), subjects are rematched every period according to their cooperativeness in the previous round; which is found to
raise cooperation levels. Likewise, if groups have a chance to exclude free-riders, this improves
cooperation in a dilemma setting (Cinyabuguma, Page et al. 2005; Croson, Fatas et al. 2008), as
does a mechanism that allows members to self-select into groups (Page, Putterman et al. 2005),
in particular if free-riders are effectively excluded by a rule that sacrifices a portion of the group
income to outsiders ( the Red Cross, as it was , Brekke, Hauge et al. 2009). Our study differs
from this literature in that we leave the incentive structure of the game completely unchanged.
All we alter are the first impressions participants happen to make – which is an option that
should naturally be available in any public good game, and in many real-life social dilemmas.
Parts of the legal literature, particularly those at the intersection of law and economics, have also
been asking how socially desirable behavior and/or compliance with the law can be brought
4
about. Again, a prominent approach is to change the incentive structure, e.g., by increasing the
expected costs of breaking the law. For example, Braga, Weisburd et al. (1999) report evidence
from a field experiment that randomly exposed 12 of 24 matched violent crime places in Jersey
City to intense police scrutiny and intervention. In the places chosen, crime rates dropped substantially, while they did not in the unaffected places. In a similar vein, in a series of sociological
field experiments, when there were signs of disorder, like graffiti, abandoned shopping carts,
littering or bicycles locked where they were not supposed to be, this induced passers-by to also
break these and other rules (Cialdini, Reno et al. 1990; Keizer, Lindenberg et al. 2008; Ramos
and Torgler 2010). In laboratory experiments, Galbiati and Vertova (2008) demonstrate that cooperation behavior increases when an explicit expectation is spelt out and enforced by (nondeterrent) sanctions (see also Kube and Traxler 2010).
Besides, our findings also underline the power of information, as it has also been observed in
other contexts. For example, previous work on voting behavior points out that information in the
form of polls (Forsythe, Myerson et al. 1993; Forsythe, Rietz et al. 1996; Klor and Winter 2007)
or of cheap-talk electoral campaigns (Corazzini, Kube et al. 2010) affect subsequent voting outcomes. Similarly, information gathered during pre-play communication2 strongly affects subjects’ decision in subsequent coordination games (Blume and Ortmann 2007), as does information from preceding asset-market outcomes (Kogan, Kwasnica et al. 2010) or information
about group members’ previous decisions (Weber 2006). Interestingly, the literature on coordination games comes to the conclusion that pre-play information promotes Nash-equilibrium play
(assuming self-centered money maximizers). By contrast, our experimental results show that the
opposite might happen in a cooperation game like the one reported here, where the socially efficient outcome is usually not part of the subgame-perfect Nash equilibrium. Pre-play information
may indeed promote socially efficient outcomes.
3
Experimental Design, Procedure, and Behavioral Predictions
Experimental Design: The baseline treatment is a typical public-good game with punishment
and counter-punishment opportunities as implemented by Nikiforakis (2008).3 The basic game
features n=4 players and consists of three stages:
At the beginning of stage 1 (“contribution stage”), players are endowed with 20 tokens each.
Players then decide simultaneously and independently how much of their endowment they want
to contribute to a public account. We denote this decision with ci . Each token that is contributed
2
3
See alsoCrawford (1998) for a general survey of experiments on communication via cheap talk.
We have chosen this design because it provides a rich environment in which different behavioral norms may
emerge. Moreover, this design has the advantage to leave sufficient room for testing the effect of both favorable and unfavorable first impressions. Had we tested a mere voluntary contribution mechanism, previous
experiments would have indicated that cooperation is very difficult to sustain in the first place. Conversely,
had we chosen a public good game with one-step punishment, successful cooperation would have been very
likely. By contrast, according to the existing literature, if we also add the counter-punishment option, expectations are in the middle between both extremes.
5
to the public account increases the payoff of each player in the group by 0.4 tokens (i.e., the
MPCR is α=.4). Each token unspent increases a player’s own payoff
by one token. The prelimi4
1
nary payoff at the end of stage 1 is thus given by π i = 20 − ci + 0.4 ch .
h =1
pij
0
1
2
3
4
5
6
7
8
9
10
C(pij)
0
1
2
4
6
9
12
16
20
25
30
Table 1: Punishment points pij per player j and associated costs C(pij) for punisher i
At the beginning of stage 2 (“punishment stage”), players are informed about every group member’s contribution to the public account. Every player i then has the opportunity to reduce the
income of each other group member j by assigning costly punishment points pij. Each punishment point received reduces a player’s income from stage 1 by ten percent. At the same time,
each punishment point assigned reduces one’s own payoff according to the cost function given in
Table 1. The preliminary payoff at the end of stage 2 is thus given by:
π i 2 = π i1


1
max 0, 10 −  p ji  −  C ( pij ) .
10
j ≠i

 j ≠i
At the beginning of stage 3 (“counter-punishment stage”), players observe who punished them
by how much in stage 2. They then have the opportunity to counter-punish the punishers by assigning them counter-punishment points cpij. The punishment technology is the same as in stage
2. Each counter-punishment point received reduces a player’s preliminary income from stage 2
by ten percent and each counter-punishment point assigned reduces one’s own payoff as given in
Table 1, but the cost of counter-punishment also depends on this player’s punishment decision.4
The final payoff at the end of stage 3 is thus given by:
π i3 = π i2


1
max 0, 10 −  cp ji  −  [C ( pij + cpij ) − C ( pij )]
10
j ≠i

 j ≠i
In treatments FAV and UNFAV, the underlying game structure remains unchanged. Compared
with the baseline the only change is that we append the experimental instructions by one additional page. The page contains selected information about unrelated groups who previously
played this game – and players know that the information is actually taken from previous experiments. They are also told that the information is selective and that all players in their group receive the same information. The information we give is the development of mean contributions
over time from four selected groups. Furthermore, we provide them with selected data about the
number of persons choosing to contribute zero in the first, resp. in the last period, and the average amount of counter-punishment that was meted out in these experiments. The way this data is
presented and the selection of the four groups are such that they constitute a positive (i.e. cooperative) impression in treatment FAV, while they constitute a negative (i.e. uncooperative) im4
That is, if i assigns counter-punishment points to j, the specific costs of the counter-punishment points depend on the number of punishment points that i has assigned to j on stage two. For example, if i had already
assigned four points to j on stage 2, assigning him a single counter-punishment point on stage three costs 3
tokens. Hence the costs of counter-punishment are given by C(pij+cpij)-C(pij).
6
pression in treatment UNFAV.5 Thus, the treatment manipulations allow us to study to what extent different contents of pre-play information affects behavior.
Procedure: The experiments were run in two different locations, namely in London (UK) and
Bonn (Germany). The baseline treatment was played in both locations. As will be seen below,
the cooperation rates in the baseline treatment differ between these two locations. The observed
cooperation rates in this environment are rather low in London, while they are rather high in
Bonn .6 Taking advantage of this observed variation between the two locations, we run treatment
FAV in London and treatment UNFAV in Bonn. Consequently we refer to the London baseline as
BASEFAV and to the Bonn baseline as BASEUNFAV. Apart from the difference in location and
thus the language of the instructions, the procedure was as similar as possible in all treatments.
All experiments were programmed using z-Tree (Fischbacher 2007). Subjects in Bonn were invited using ORSEE (Greiner 2004). The composition of subjects with respect to age, gender and
field of studies was similar between locations and between treatments. Subjects were not allowed
to participate in more than one treatment (between-subject design). When subjects arrived in the
lab, they were seated in separate cubicles. The experimental instructions were then handed out to
them and read out aloud. This was done to create common knowledge and to ensure that everybody had read and understood the instructions. Additionally, subjects had to take a quiz and
could pose comprehension questions in private before the game started. The instructions were
written in neutral language, avoiding potentially loaded terms like punishment or public good
(cp. Appendix 1).
After subjects had finished reading the instructions, in treatments FAV and UNFAV they were
provided with the additional information sheet (cp. Appendix 3 and 4). Subjects were then randomly divided into groups of four and played the above described game repeatedly for ten consecutive periods. A partner protocol was used, i.e., the group composition stayed constant over
the entire 10 periods of anonymous interaction. In the end, subjects were privately paid their cumulated earnings and left.
We had 60 participants (3 sessions) in treatment FAV in the Laboratory for Experimental Economics at the University of Bonn in 2009. Participants received their accumulated earnings from
the experiment (1 token = Euro 0.04) and an additional show-up fee of 5 Euro. On average, a
session lasted 60 minutes and subjects earned Euro 14.46. The four sessions for treatment UNFAV were run in 2010 at the experimental laboratory of Royal Holloway, University of London,
UK. We used the same lab and recruited 64 students from the same subject pool as in Nikiforakis
(2008). We also used the same software. Again, participants received their accumulated earnings
from the experiment (1 token = 0.04 British Pounds). On average, a session lasted 60 minutes
5
6
The Appendix contains the exact wording of these two pages including the graphs, as well as a translation of
the general instructions that were used in all treatments.
The corresponding data that we use for our baseline treatments BASEFAV and BASEUNFAVare taken from
Nikiforakis (2008) and Beckenkamp, Engel et al. (2009), respectively. The data reported from treatments
FAV and UNFAV were collected by us. They are novel and not reported elsewhere. We are thankful to Nikos
Nikiforakis for providing us with his data.
7
and subjects earned 9.06 Pounds. In the two baseline treatments, analogous procedures had been
used. The datasets comprise 48 subjects in BASEFAV and 68 participants in BASEUNFAV. Additional details on the two baselines can be found in Nikiforakis (2008) and Beckenkamp, Engel et
al. (2009), respectively.
Behavioral Predictions: For rational, self-centered money maximizing players, the game at
hand is a cooperation problem. Using backwards induction, it is straightforward that the unique
subgame perfect equilibrium is to contribute nothing to the public good at the first stage, and not
to punish nor counter-punish at subsequent stages. This leads to an equilibrium payoff of 20 tokens per period and player. By contrast, the socially efficient outcome would be achieved if everyone contributed their entire endowment, in which case every player would earn 32 tokens per
period. But in that case, each player would have an individual incentive to free-ride on the others’ contributions, which would yield him a payoff of 44 tokens, ceteris paribus. Since we announce the number of periods, through unraveling this is also the prediction for the repeated
game. The standard prediction of no cooperation not only holds true in the baseline treatment.
Providing subjects with additional information about other groups does not change the equilibrium prediction. Therefore, with rational, self-centered money maximizing players we should not
expect to observe different behavior in treatments FAV or UNFAV.
By contrast, things might change as soon as we allow for social preferences, in particular if we
expect a substantial fraction of the population to act as conditional cooperators. If conditional
cooperators are sufficiently optimistic about the cooperativeness of their interaction partners,
cooperation gives them higher utility than defection. Facing a public good, conditionally cooperative subjects should therefore base their decision on what they believe other subjects to do. The
beliefs are likely to be related to other group members’ behavior in the previous period(s). Yet,
upon the first encounter (in the first period), a conditional cooperator needs to form initial expectations of how others are likely to behave. This is where our treatment manipulation might make
a difference. Pre-play information about other, unrelated groups might influence a participant in
forming initial beliefs, which should then guide her first-period decision. If her (conditionally
cooperative) group peers were also influenced by the selective pre-play information, then group
contributions in the first period are likely to mirror the pre-play information. Her beliefs for the
second period will thus be very similar again, and so will contributions be in the second period;
and similarly in subsequent periods. In that case, we would expect to observe more cooperation
in treatment FAV than in BASEFAV, and less cooperation in treatment UNFAV than in BASEUNFAV throughout the game.
4
Results
In the following, we will first briefly describe the results from the two baseline treatments BASEFAV and BASEUNFAV. As will be seen, observed cooperation rates are low in London while they
are high in Bonn. Afterwards, we move on to show that the treatment manipulations have a sub8
stantial and significant influence on subject’s behavior. Compared to the respective baseline
treatment, cooperation rates are lower in treatment UNFAV while they are higher in treatment
FAV. As will be seen, this is likely to be driven by the change in first period’s contributions
which, in turn, are sensitive to our manipulation of first impressions. In line with this, group’s
behavior in the first period is strongly correlated with its future development, as we subsequently
show. In the last subsection, additional experiments further underline our result by explicitly exploring how beliefs react to the treatment manipulations.
BASELINES:
Figure 1 shows the mean development of contributions over time in BASEFAV and BASEUNFAV.
As can be seen, the environment in which subjects interact is complex and allows for different
behavioral patterns.7 Without any manipulation, cooperation rates are rather low in London
(BASEFAV). The mean contribution in BASEFAV is 8.72. By contrast, subjects in Bonn (BASEUNFAV) display a high degree of cooperation. They contribute on average 16.45 to the public good.
We therefore test in Bonn whether poor first impressions reduce the tendency to contribute
much, and thereby trigger a vicious cycle. Conversely we test in London whether promising first
impressions increase the willingness to give more initially, and thereby trigger a virtuous cycle.
Contributions to PG
20
18
16
14
12
10
8
6
4
2
1
2
3
4
5
6
Period
BASE FAV
7
8
9
10
BASE UNFAV
Fig. 1 – Mean Contribution over time in BASEFAV and BASEUNFAV
Notes: With 95% Confidence intervals around the period means. Group clustered standard errors.
BASEFAV(Nikiforakis 2008) has 48 subjects (12 groups). BASEUNFAV(Beckenkamp, Engel et al. 2009)
has 68 subjects (17 groups).
7
While this observation, in particular the behavioral differences between locations, is interesting per se, here
we do not want to elaborate on this point but would like to refer the interested reader to (Beckenkamp, Engel
et al. 2009) for an extensive discussion.
9
BREAKING WINDOWS:
What happens if we provide subjects with additional information that conveys a negative preplay impression? As can be seen in Figure 2, this experimental manipulation has a substantial
effect on contributions. After “breaking the window” with an “experimental hammer”, behavior
changes radically. Whereas in BASEUNFAV subjects choose full contribution in 60% (411/680) of
all cases, in UNFAV they only do so in 20% (120/600) of all cases. Also at the opposite end of the
contribution scale the effect is evident. In BASEUNFAV subjects only in 8.5% (58/680) of all cases contribute less than 5 tokens. With the treatment manipulation, this fraction grows to almost
32% (191/600). Overall, the mean contribution drops significantly from 16.45 in BASEUNFAV
down to 10.71 in UNFAV (Mann Whitney ranksum test, N = 32, p=0.0024, two-sided). Hence,
our “experimental hammer” indeed creates uncooperative behavior in an otherwise cooperative
environment, just like Zimbardo’s sledgehammer generated Bronx-like behavior in Palo Alto.
Moreover, the effect is long-lasting: the initial manipulation of expectations still leads to considerably lower cooperation even after ten periods of interaction, i.e. after 9 possible opportunities
to revise expectations.
20
Contributions to PG
18
16
14
12
10
8
6
1
2
3
4
5
6
Period
BASE UNFAV
7
8
9
10
UNFAV
Fig. 2 – Mean Contribution over time in BASEUNFAV and UNFAV
Notes: With 95% Confidence intervals around the period means. Group clustered standard errors.
BASEUNFAV (Beckenkamp, Engel et al. 2009) has 68 subjects (17 groups). UNFAV has 60 subjects (15
groups).
10
FIXING WINDOWS:
From a policy perspective, the opposite situation is even more interesting. Is the mere fact that
one “repairs windows” enough to foster cooperation? To test this hypothesis, in treatment FAV
we provide participants with favorable first impressions. Inspecting Figure 3, it is plain that repairing windows is less powerful than breaking them. Descriptively, once participants have seen
graphs from groups that have cooperated successfully, they contribute more. Average contributions increase from 8.7 in BASEFAV to 12 in FAV. This increase of almost 40% is remarkable,
though it falls short of being statistically significant (ranksum test, p=.1567, two-sided). Still, the
behavioral pattern seems to be strongly influenced by the favorable manipulation. While cooperation is sustained in FAV, contribution rates quickly decline in BASEFAV. Consequently, from the
second period on, mean contributions in BASEFAV are outside the 95% confidence interval of
mean contributions in FAV after participants have seen graphs from cooperative groups. Contributions rise sharply for the first three periods. They then stay more or less stable until period six.
There is a slight decay for the final periods. Compare this to the data from treatment BASEFAV.
Contributions never rise. They decay from the fourth period on. This difference of trends is significant (Mann Whitney over means of first differences, N = 28, p = .0243). Our data thus suggests that fixing windows pays. The beneficial effect only takes longer to unfold.
16
Contributions to PG
14
12
10
8
6
4
2
1
2
3
4
5
6
Period
BASE FAV
7
8
9
10
FAV
Fig. 3 – Mean Contribution over time in BASEFAV and FAV
Notes: With 95% Confidence intervals around the period means. Group clustered standard errors.
BASEFAV(Nikiforakis 2008) has 48 subjects (12 groups). FAV has 64 subjects (16 groups).
11
INITIAL CONTRIBUTIONS AND FUTURE DEVELOPMENT:
How does this very simple and light-handed mechanism of providing subjects with information
from unrelated groups create these strong differences in the overall cooperation rate? If we take a
closer look at our data, the answer to this question seems to be twofold.
Mean contributions in period 1
20
15
10
5
0
0
5
10
15
Mean contributions in later periods
BASE FAV
FAV
20
BASE UNFAV
UNFAV
Fig. 4 – Initial contributions vs. future group development
Notes: Every entry in the graph represents one specific 4-person group. There are 12 groups in
BASEFAV, 17 in BASEUNFAV, 16 in FAV and 17 in UNFAV. “Later periods” refers to the mean contribution in periods 2 to 10. The 45 degree line depicts all points in which the average contribution in period 1 is identical to the average contribution in later periods.
First, we find that in all treatments contribution behavior in period 1 significantly predetermines
cooperativeness in later rounds. This can best be seen when we look at the development in individual groups. Figure 4 compares the average contribution to the public good in period 1 with the
average contribution in periods 2 to 10. There is a clear correlation between initial and future
cooperation. The correlation is strong and significant, over all treatments (Spearman’s
rho=.6981, N = 60, p<.001) as well as for each treatment separately.
12
Share of subjects (in %)
BASE FAV
BASE UNFAV
40
40
30
30
20
20
10
10
0
0
0
5
10
FAV
15
20
40
40
30
30
20
20
10
10
0
0
0
5
10
0
5
15
20
0
5
Initial contribution to PG
10
15
UNFAV
10
15
20
20
Fig. 5 – Initial contributions to the public good
Notes:BASEFAV has 48 subjects, BASEUNFAV 68, FAV 64, and UNFAV 60.
Second, we observe that our experimental manipulation has a substantial effect on initial contributions to the public good (cp. Figure 5). In BASEUNFAV, 40% of subjects go for full contribution in the first period. Yet if they receive the unfavorable information in treatment UNFAV, only
16% do so. The mean contribution in the first period drops by almost 30%, from 13.64 in BASEUNFAV down to 9.78 in UNFAV; the difference being significant (ranksum test, p=.0007, twosided). Our favorable manipulation does not have the same power. Our data suggest that “fixing
windows” in an uncooperative environment is not as straightforward as breaking them in an otherwise cooperative environment. The favorable information increases the mean contribution in
the first period only very slightly, from 10.35 in BASEFAV to 10.48 in FAV. This difference is
not significant.
Taken together, these two findings provide a possible explanation for the overall difference in
cooperation that we observe between treatments. Contributions are path-dependent, i.e., they depend on contributions in the first period. These first contributions, in turn, are apparently manipulable by our very simple mechanism (though the strength of this manipulation seems to be
13
asymmetric). Thus, if people are provided with unfavorable information, i.e., if they observe
“broken windows”, they start with low cooperation rates and also stay at low cooperation rates in
the long run; favorable information at least makes it easier to sustain cooperation in the long run.
This suggests that subjects seem to form an initial expectation about how others are likely to behave and base their decision on this expectation. The initial expectations then appear as some
kind of self-fulfilling prophecy: The mere pre-game expectation of an uncooperative environment leads to an irrevocable loss of cooperation throughout the game (and vice versa).
PRE-GAME EXPECTATIONS:
To shed more light on the role of expectations, we subsequently ran additional experiments in
2010 at the Econ Lab of the University of Jena, Germany. In these experiments, subjects received exactly the same instructions and had to answer the same control questions as in the
games above. However, instead of actually playing the game they were then told that these experiments had been conducted before. Their task was to guess how much those previous participants contributed on average to the common project in the first period (rounded to the next integer). Moreover, subjects had to state how confident they were in this expectation (on a Likert
scale ranging from 1 to 7, where 1 means “being very unsure” and 7 means “being very sure”
about the belief).8
BASE
25
FAV
6
20
UNFAV
25
6
2 5
6
20
2 0
5
5
15
5
15
1 5
4
4
10
5
0
4
10
1 0
3
3
3
5
5
2
2
2
0
5
10
15
20
Confidence
Share of subjects (in %)
Altogether, 96 subjects participated in these experiments. In the first treatment, subjects received
the instructions from the baselines. In the second treatment, they received the instructions of
treatment Fav. Finally in the third treatment they received the instructions of treatment Unfav.
0
0
0
5
10
15
20 0
5
10
15
20
Belief about initial mean contribution to PG
Fig. 6 –Beliefs and Confidence
Notes:BASE has 48 subjects, FAV and UNFAV have 24 subjects respectively. The bars denote the share of subjects
in a given treatment expressing a certain belief. The dashed lines represent linear prediction plots for the subjects’
self-reported Confidence (with 95% confidence intervals). Confidence is measured on a 7-point Likert scale ranging
from 1 (“I am very unsure”) to 7 (“I am very sure”).
8
In the belief elicitation stage, subjects had to guess a whole number between 0 and 20. If the guess was exactly correct, they received 4 Euros. If the guess was wrong, they received 0 Euros. The confidence question
was not incentivized. See Appendix 5 for details.
14
As Fig. 6 shows, the information manipulation has a strong effect on beliefs. In BASE, subjects
expect the first period’s contribution to be 10.27 on average. The belief drops to 8.04 in UNFAV,
i.e. if they know that participants received the unfavorable information sheet. The difference is
significant (ranksum test, p=0.0472, two-sided). By contrast, the average belief increases from
10.27 to 12.96 in FAV. Again, the difference is significant (ranksum test, p=0.0054, two-sided).
Mean confidence does not vary significantly across treatments. Interestingly however, as can be
seen in Fig. 6, in each treatment confidence is negatively correlated with beliefs. Subjects with a
relatively pessimistic view of other’s cooperativeness are considerably more confident than those
with optimistic beliefs. This pattern is significant and very robust9 and might give us an indication for why actual contributions react strongly to negative selective information, but only cautiously to positive selective information. The less confident optimistic beliefs might be more
prone to external manipulation than the more confident pessimistic beliefs.
Taken together, the additional experiments provide further evidence for the above interpretation
of our results from the repeated public-good game. If people face a cooperation problem and observe “broken windows”, they strongly expect that others are not likely to cooperate, so they
themselves also do not cooperate. By contrast, if “windows are fixed”, they believe that others
might cooperate, but at the same time acknowledge the cooperation dilemma. Thus, they are not
very confident in their expectation. This translates into only intermediate rates of cooperation –
though this seemingly suffices to at least sustain cooperation over time.
5
Discussion
In this paper, we have explored whether cooperative behavior reacts to selective pre-play information about other, unrelated groups. To this end, we have used an experimental framework that
captures cooperation situations which are sufficiently rich, so that several potentially conflicting
behavioral norms might emerge. We find that the aggregate level of cooperativeness is dramatically reduced by giving subjects examples of uncooperative behavior. The opposite intervention
was less powerful, but still effective. If participants in an otherwise uncooperative environment
saw examples of successful groups, cooperation rates were stabilized, while they quickly decayed otherwise. A likely reason for these findings is that cooperation is path-dependent, i.e.,
long-run behavior is strongly correlated with initial contribution rates – which, in turn, are a reaction to the pre-play information given to the subjects. The latter might be due to a change in
subjects’ expectations about others’ behavior, as was suggested by our additional experiments.
Our findings have a number of important implications. First, they clearly point to the relevance
of pre-game communication – a factor which has only lately started to receive significant attention in the literature. While the existing literature usually focuses on self-chosen cheap talk mes9
Log-Level OLS regression of ln(confidence) on belief. A 1-point increase in beliefs corresponds on average
to a 2 % decrease in confidence (p=0.004). This result is robust to adding treatment dummies, changing the
functional form to Level-Level or Log-Log, or running a Tobit.
15
sages (for an overview, see Crawford 1998), we demonstrate that also exogenously selected,
one-way information about other players can alter how players act in subsequent games. In particular the findings of our second series of experiments might be of interest to this literature.
They suggest a possible channel through which the observed effects of cheap talk are mediated,
namely through the alteration of subjects’ pre-game expectations.
Second, and closely related, is our finding that there seems to be a strong asymmetry in the
strength of the reaction to the mechanism. By giving subjects examples of uncooperative behavior, the aggregate level of cooperativeness is immediately reduced. By contrast, optimistic foreign experiences need time to unfold their beneficial effect. Bad impressions carry more weight
than good ones, so that creating bad behavior is seemingly easier than producing good one – a
finding which is also mirrored in previous work on the prevalence of “bad” vs. “good” (cf.
Baumeister, Bratslavsky et al. 2001). Thus, one might expect that also other mechanisms which
build on the manipulation of beliefs perform differently, depending on the content of the experience as well as on the specific environment in which they occur.
Taken together, our results underline the power and importance of information and experience in
shaping cooperative behavior. The bottom line is that observation matters. Interestingly, people
do not only learn from what they experience themselves. They also seem to learn “vicariously”,
by observing others, or by seeing the results (Bandura 1977). The effect is even present when
participants are told that the information they are receiving is selective. This connects our study
with the growing literature on social learning (for a recent meta-study see Weizsäcker 2010). Our
subjects’ behavior critically depends on pre-game expectations, which we show to be easy to
deteriorate in a complex setting – simply by providing subjects the opportunity of vicarious
learning.
This suggests that, in appropriate circumstances, impression management might indeed be a feasible tool to avert, or at least to mitigate, the danger of social dilemmas. Of course this is a paternalistic intervention. But note that conditional cooperators need not even be deceived for the intervention to be successful. All that is required is that they expect the manipulation to matter for
a sufficient fraction of the remaining members of their group, be that because they are deceived,
or because they are skeptical themselves, but willing to give cooperation a try since the intervention gives all of them one and the same informational starting point.
For policy makers these findings represent both a chance and a peril. If they do not manage to
repair broken windows quickly, both literally and metaphorically speaking, chances are a vicious
cycle starts. By contrast, if they can induce some to lead others by their socially beneficial example, this strategy may well work. In particular, policy makers would want to prevent (perhaps
wrong) pessimistic beliefs from spreading. In any case, we show that home-grown expectations
must not be disregarded in order to attain socially desirable outcomes. Consequently in situations
where the success of a law depends on the willingness of individual citizens to cooperate – for
instance in areas like waste separation and sustainable water use – government might want to
16
consider a PR campaign in order to create a general atmosphere of cooperativeness within the
population.
Managing first impressions might certainly be less effective if the large majority of addressees
know better. The intervention requires a sufficient degree of uncertainty. Yet in political reality,
quite a few public goods are characterized by deep conceptual and factual uncertainty. Problems
like climate change are heavily contested among scientists and not well understood by many.
Addressees have to trust expertise. If in the eyes of addressees the underlying social problem is
opaque, they are also likely to be uncertain how others will react to it. Most importantly, addressees face behavioral uncertainty whenever they newly enter a community. They do not know
local mores, nor do they know how determined the group is when it comes to enforcing them.
The results in this paper should, of course, not be taken as arguments against the importance of
elaborated, incentives-altering mechanisms in general. For instance, the effectiveness of our very
simple mechanism seems to be asymmetric, since it has a hard time to increase cooperation in an
otherwise uncooperative environment. Still, our findings suggest that minimal interventions can
have a strong behavioral effect. Future research could try to explore the interaction between such
simple and other, more complex and intrusive mechanisms. Besides, it might also be informative
to study in more detail if certain player “types” (like selfish actors, conditional cooperators, or
altruists) are affected differently by mechanisms in the spirit of the one presented here. It is even
conceivable that such interventions change the preferences of certain players. Another important,
though challenging, next step would certainly be to move into the field and to investigate the
effectiveness of means to foster cooperation in a more natural setup. Extending on our work allows for studying these and other interesting aspects in the future.
17
References
ANDREONI, JAMES (1988). "Why Free Ride? Strategies and Learning in Public Goods Experiments." Journal of Public Economics 37: 291-304.
ANDREONI, JAMES and RAGAN PETRIE (2004). "Public Goods Experiments without Confidentiality. A Glimpse into Fund-raising." Journal of Public Economics 88: 1605-1623.
BANDURA, ALBERT (1977). Social Learning Theory. Englewood Cliffs, Prenctice Hall.
BAUMEISTER, ROY, ELLEN BRATSLAVSKY, CATRIN FINKENAUER and KATHLEEN D. VOHS (2001).
"Bad is Stronger than Good." Review of General Psychology 5: 323-370.
BECKENKAMP, MARTIN, CHRISTOPH ENGEL, ANDREAS GLÖCKNER, BERND IRLENBUSCH, HEIKE
HENNIG-SCHMIDT, SEBASTIAN KUBE, MICHAEL KURSCHILGEN, ALEXANDER MORELL, ANDREAS NICKLISCH, HANS-THEO NORMANN and EMANUEL TOWFIGH (2009). Beware of
Broken Windows! First Impressions in Public-good Experiments, mimeo.
BLUME, ANDREAS and ANDREAS ORTMANN (2007). "The Effects of Costless Pre-Play Communication. Experimental Evidence from Games with Pareto-Ranked Equilibria." Journal of
Economic Theory 132: 274-290.
BRAGA, ANTHONY A., DAVID L. WEISBURD, ELIN J. WARING, LORRAINE GREEN MAZEROLLE,
WILLIAM SPELMAN and FRANCIS GAJEWSKI (1999). "Problem-Oriented Policing in Violent
Crime Places. A Randomized Controlled Experiment." Criminology 37: 541-580.
BREKKE, KJELL ARNE, KAREN EVELYN HAUGE, JO THORI LIND and KARINE NYBORG (2009).
Playing with the Good Guys: A Public Good Game with Endogenous Group Formation,
mimeo.
BURLANDO, ROBERTO M. and FRANCESCO GUALA (2005). "Heterogeneous Agents in Publilc
Goods Experiments." Experimental Economics 8: 35-54.
CAMERON, ADRIAN COLIN and P. K. TRIVEDI (2005). Microeconometrics. Methods and Applications. Cambridge ; New York, Cambridge University Press.
CIALDINI, ROBERT B., RAYMOND R. RENO and CARL A. KALLGREN (1990). "A Focus Theory of
Normative Conduct. Recycling the Concept of Norms to Reduce Littering in Public Places." Journal of Personality and Social Psychology 58: 1015-1026.
CINYABUGUMA, MATTHIAS, TALBOT PAGE and LOUIS PUTTERMAN (2005). "Cooperation under
the Threat of Expulsion in a Public Goods Experiment." Journal of Public Economics 89:
1421-1435.
CORAZZINI, LUCCA, SEBASTIAN KUBE, MICHEL ANDRÉ MARÉCHAL and ANTONIO NICOLO (2010).
Elections and Deceptions. Theory and Experimental Evidence
18
CRAWFORD, VINCENT (1998). "A Survey of Experiments on Communication via Cheap Talk."
Journal of Economic Theory 78(2): 286-298.
CROSON, RACHEL T.A., ENRIQUE FATAS and TIBOR NEUGEBAUER (2008). The Effect of Excludability on Team Production, mimeo.
DUFWENBERG, MARTIN and GEORG KIRCHSTEIGER (2004). "A Theory of Sequential Reciprocity." Games and Economic Behavior 47: 268-298.
FALK, ARMIN and URS FISCHBACHER (2006). "A Theory of Reciprocity." Games and Economic
Behavior 54: 293-315.
FALKINGER, JOSEF, ERNST FEHR, SIMON GÄCHTER and RUDOLF WINTER-EBMER (2000). "A Simple Mechanism for the Efficient Provision of Public Goods. Experimental Evidence."
American Economic Review 90: 247-264.
FEHR, ERNST and SIMON GÄCHTER (2000). "Cooperation and Punishment in Public Goods Experiments." American Economic Review 90: 980-994.
FEHR, ERNST and SIMON GÄCHTER (2002). "Altruistic Punishment in Humans." Nature 415: 137140.
FISCHBACHER, URS (2007). "z-Tree. Zurich Toolbox for Ready-made Economic Experiments."
Experimental Economics 10: 171-178.
FISCHBACHER, URS and SIMON GÄCHTER (2010). "Social Preferences, Beliefs, and the Dynamics
of Free Riding in Public Good Experiments." American Economic Review 100: 541-556.
FISCHBACHER, URS, SIMON GÄCHTER and ERNST FEHR (2001). "Are People Conditionally Cooperative? Evidence from a Public Goods Experiment." Economics Letters 71: 397-404.
FORSYTHE, ROBERT, ROGER B. MYERSON, THOMAS RIETZ and ROBERT J. WEBER (1993). "An
Experiment on Coordination in Multi-Candidate Elections. The Importance of Polls and
Election Histories." Social Choice & Welfare 10: 223-247.
FORSYTHE, ROBERT, THOMAS RIETZ, ROGER B. MYERSON and ROBERT J. WEBER (1996). "An
Experimental Study of Voting Rules and Polls in Three-Candidate Elections." International
Journal of Game Theory 25: 355-383.
GÄCHTER, SIMON and CHRISTIAN THÖNI (2007). "Social Learning and Voluntary Cooperation
Among Like-Minded People." Journal of the European Economic Association 3: 303-314.
GALBIATI, ROBERTO and PIETRO VERTOVA (2008). "Law and Behaviours in Social Dilemmas.
Testing the Effect of Obligations on Commitment." Games and Economic Behavior 64:
146-170.
19
GLÖCKNER, ANDREAS, BERND IRLENBUSCH, ANDREAS NICKLISCH and HANS-THEO NORMANN
(2010). "Leading With(out) Sacrifice? A Public-Goods Experiment with a Privileged Player." Economic Inquiry ***: ***.
GOERG, SEBASTIAN J. and GARY WALKOWITZ (2010). "On the Prevalence of Framing Effects
Across Subject-Pools in a Two-Person Cooperation Game." Journal of Economic Psychology ***: ***.
GREINER, BEN (2004). An Online Recruiting System for Economic Experiments. Forschung und
wissenschaftliches Rechnen 2003. Kurt Kremer und Volker Macho. Göttingen: 79-93.
GUNNTHORSDOTIR, ANNA, DANIEL HOUSER and KEVIN MCCABE (2007). "Disposition, History
and Contributions in Public Goods Experiments." Journal of Economic Behavior & Organization 62: 304-315.
GÜRERK, ÖZGÜR, BERND IRLENBUSCH and BETTINA ROCKENBACH (2006). "The Competitive
Advantage of Sanctioning Institutions." Science 312: 108-111.
HERRMANN, BENEDIKT, CHRISTIAN THÖNI and SIMON GÄCHTER (2008). "Antisocial Punishment
Across Societies." Science 319: 1362-1367.
ISAAC, R. MARK, JAMES M. WALKER and ARLINGTON W. WILLIAMS (1994). "Group Size and the
Voluntary Provision of Public Goods. Experimental Evidence Utilizing Large Groups."
Journal of Public Economics 54: 1-36.
KEIZER, KEES, SIEGWART LINDENBERG and LINDA STEG (2008). The Spreading of Disorder. Science. 322: 1681-1685.
KLOR, ESTEBAN and EYAL WINTER (2007). "The Welfare Effects of Public Opinion Polls." International Journal of Game Theory 35: 379-394.
KOGAN, SHIMON, ANTHONY M. KWASNICA and ROBERTO A. WEBER (2010). "Coordination in the
Presence of Asset Markets." American Economic Review ***: ***.
KUBE, SEBASTIAN and CHRISTIAN TRAXLER (2010). "The Interaction of Legal and Social Norm
Enforcement." Journal of Public Economic Theory ***: ***.
MAMMEN, ENNO (1993). "Bootstrap and Wild Bootstrap for High Dimensional Linear Models."
Annals of Statistics 21: 255-285.
MENGEL, FRIEDRIKE (2007). "The Evolution of Function-Value Traits for Conditional Cooperation." Journal of Theoretical Biology 245: 564-575.
NIKIFORAKIS, NIKOS S. (2008). "Punishment and Counter-Punishment in Public Good Games:
Can We Really Govern Ourselves?" Journal of Public Economics 92: 91-112.
20
OSTROM, ELINOR, JAMES M. WALKER and ROY GARDNER (1992). "Covenants with and without
Sword. Self-Governance is Possible." American Political Science Review 40: 309-317.
PAGE, TALBOT, LOUIS PUTTERMAN and BULENT UNEL (2005). "Voluntary Association in Public
Goods Experiments. Reciprocity, Mimicry and Efficiency." Economic Journal 115: 10321053.
RAMOS, JOAO and BENNO TORGLER (2010). Are Academics Messy? Testing the Broken Windows Theory with a Field Experiment in the Work Environment
http://www.bepress.com/cgi/viewcontent.cgi?article=1505&context=feem.
WEBER, ROBERTO A. (2006). "Managing Growth to Achieve Efficient Coordination in Large
Groups." American Economic Review 96: 114-126.
WEIZSÄCKER, GEORG (2010). "Do We Follow Others When We Should? A Simple Test of Rational Expectations." American Economic Review ***.
WILSON, JAMES Q. and GEORGE L. KELLING (1982). "Police and Neighborhood Safety. Broken
Windows." Atlantic Monthly 127: 29-38.
WU, C.F.J. (1986). "Jackknife, Bootstrap and Other Resampling Methods in Regression Analysis." Annals of Statistics 14: 1261-1295.
ZIMBARDO, PHILIP (1969). "The Human Choice. Individuation, Reason, and Order versus Deindividuation, Impulse, and Chaos." Nebraska Symposium on Motivation 17: 237-307.
ZIMBARDO, PHILIP G. and EBBE B. EBBESEN (1969). Influencing Attitudes and Changing Behavior. A Basic Introduction to Relevant Methodology, Theory, and Applications. Reading,
Mass.,, Addison-Wesley Pub. Co.
21
Appendix 1: Instructions Baseline
You are now taking part in an economic experiment. If you read the following instructions carefully, you can, depending on your decisions, earn a considerable amount of money. It is therefore important that you take your time to understand the instructions.
The instructions which we have distributed to you are for your private information. Please do not communicate with
the other participants during the experiment. Should you have any questions please ask us.
During the experiment we shall not speak of Pounds, but of Experimental Currency Units (ECU). Your entire earnings
will be calculated in ECUs. At the end of the experiment the total amount of ECUs you have earned will be converted
to Pounds at the rate of 1 ECU = 4 p and will be immediately paid to you in cash.
At the beginning of the experiment the participants will be randomly divided into groups of four. You will therefore be
in a group with 3 other participants. The composition of each group will remain the same throughout the experiment. The experiment lasts 10 periods and each period is divided into 3 stages.
The 1st stage:
At the beginning of each of the 10 periods each participant will receive 20 ECUs. In the following, we shall refer to
this amount as the “endowment”. In the 1st stage, your task is to decide how to use your endowment. You have to
decide how many of the 20 ECUs you want to contribute to a project (from 0 to 20) and how many of them to keep for
yourself. The consequences of your decision are explained in detail below.
Once all the players have decided their contribution to the project you will be informed about, the group’s total contribution, your income from the project and your payoff in this period. Your payoff in each period is calculated using the
following simple formula. Again, if you have any difficulties do not hesitate to ask us.
Income from the 1st
stage
= Endowment of
ECUs
– Your contribution to the
Project
+ 0.4*Total contribution to the
Project
st
This formula shows that your 1 stage income consists of two parts:
1) The ECUs which you have kept for yourself (endowment – contribution)
2) The income from the project, which equals to the 40% of the group’s total contribution.
The income of each group member from the project is calculated in the same way. This means that each group
member receives the same income from the project. Suppose the sum of the contributions of all group members are
60 ECUs. In this case, each member of the group receives an income from the project of: 0.4*60=24 ECUs. If the
total contribution to the project is 9 points, then each member of the group receives an income of: 0.4*9=3.6 ECUs
from the project.
You always have the option of keeping the ECUs for yourself or contributing them to the project. Each ECU that you
keep raises your end of period income by 1 ECU. Supposing you contributed this point to the project instead, then the
total contribution to the project would rise by 1 ECUs. Your income from the project would thus rise by 0.4*1=0.4
ECUs. However, the income of the other group members would also rise by 0.4 ECUs each, so that the total income
of the group from the project would be 1.6 points. Your contribution to the project therefore also raises the income of
the other group members. On the other hand you also earn an income for each point contributed by the other members to the project. In particular, for each point contributed by any member you earn 0.4 ECUs.
In addition to the 20 ECUs per period, each participant receives a one-off lump sum payment of 25 ECUs at the beginning of this part. This one-off payment can be used to pay for eventual losses during the experiment. However,
you can always evade losses with certainty through your own decisions. Note that this lump sum payment will
not be used to calculate the income from the period. It will only be added to your total income from all the periods at
the very end.
The 2nd stage:
At the 2nd stage you will be informed how much each group member contributed individually to the project at the 1st
stage. At this stage you can reduce or leave equal the income of each member of your group by distributing
points. The other group members can also reduce your income if they wish to.
If you choose 0 points for a particular group member, you do not change his or her income. However if you give a
member 1 point you reduce his or her income by 10 percent. If you give a member 2 points you reduce his or her
income by 20 percent, etc. The amount of points you distribute to each member determines, therefore, how much you
st
reduce their income from the 1 stage. If one player receives in total 4 points his income will be reduced by 40% and
st
if he receives 10 or more his income from the 1 stage will be reduced by 100%.
22
If you distribute points you have costs in ECUs, which depend on the amount of points you distribute. You can distribute between 0 and 10 points to each group member. The more points you give to any group member, the higher
your costs. Your total costs are equal to the sum of the costs of distributing points to each of the other three
group members. The following table illustrates the relation between distributed points to each group member and
the cost of doing so in ECUs.
Points
Cost of points per
person
0
1
2
3
4
5
6
7
8
9
10
0
1
2
4
6
9
12
16
20
25
30
Example: Supposing you give 2 points to player 1 this costs you 2 ECUs; if you also give 8 points to player 3 this
costs you a further 20 ECUs; and if you give 0 points to the last group member this has no additional cost for you. In
this case, your total costs of distributing points would be 22 ECUs (2+20+0) and not 30 ECUs.
The following equation summarizes the previous information. Your total income from the two stages is calculated as
follows:
= Income from the 1st
stage
Income at the end
nd
of the 2 stage
* [(10 – received points)/10]
– Costs of distributed points
nd
rd
Please note that your income in ECUs at the end of the 2 and the 3 stage can be negative, if the costs of your
st
points distributed exceeds your (possibly reduced) income from the 1 stage. You can however evade such losses
nd
with certainty through your own decisions. Should your income become zero or negative at the end of the 2
rd
rd
stage you will not be able to continue to the 3 stage. If your income becomes zero or negative at the end of the 3
stage you can simply use your 25 ECUs that we gave you in the beginning in order to pay this off.
The 3rd stage:
In the 3rd and final stage, after being informed of the points that the other group members assigned to you, you will be
given one last opportunity of assigning points back to the other participants, thus reducing their income. We shall call
these points “counter-points”. You will only be able to assign counter-points to participants who assigned
points to you during the 2nd stage
The costs of assigning points, as well as the income reduction caused by each point remain the same as before. The
following table shows you how to calculate the costs for assigning counter-points.
Costs for assigning counter-points to one specific group member in the
3rd stage
Points that you already assigned to
nd
one specific group member in the 2
Counter-points you assign to that same group member in the 3rd stage
stage
0
1
2
3
4
5
6
7
8
9
10
0
0
1
2
4
6
9
12
16
20
25
30
1
0
1
3
5
8
11
15
19
24
29
2
0
2
4
7
10
14
18
23
28
3
0
2
5
8
12
16
21
26
4
0
3
6
10
14
19
24
5
0
3
7
11
16
21
6
0
4
8
13
18
7
0
4
9
14
8
0
5
10
9
0
5
10
0
nd
rd
Example: if you distribute 2 points in the 2 stage to player 1 you have a cost of 2 ECUs If in the 3 stage you decide
to distribute 3 counter-points to player 1, a further 7 ECUs are added to your cost.
23
Your income after the 3rd stage (= period income) is therefore calculated as follows:
Income at the end
rd
of the 3 stage
= Income from the 2nd * [(10 – received counterstage
points)/10]
– Cost of distributed counterpoints
If you have any further questions please raise your hand and one of the supervisors
24
Appendix 2: Control Questionnaire
1.
Each group member has an endowment of 20 ECUs. Nobody (including you) contributes any ECUs to the project. What is:
2.
a.
Your income at the end of the first stage? ………
b.
The income of the other group members?………
Each group member has an endowment of 20 ECUs. You contribute 20 ECUs to the project. All other group
members contribute 20 ECUs each to the project. What is:
3.
a.
Your income at the end of the first stage? ………
b.
The income of the other group members?………
Each group member has an endowment of 20 ECUs. The other three group members contribute together a total
of 30 ECUs to the project. What is:
4.
a.
Your income at the end of the first stage if you contribute 0 ECUs to the project? ……….
b.
Your income at the end of the first stage if you contribute 15 ECUs to the project? ……….
Each group member has an endowment of 20 ECUs. You contribute 8 ECUs to the project. What is:
a.
Your income at the end of the first stage if the other group members together contribute a further total of 7
b.
Your income at the end of the first stage if the other group members together contribute a further total of 22
ECUs to the project?...............
ECUs to the project?...............
5.
At the second stage you distribute the following points to your three other group members: 9, 5, 0. What are the
total costs of your distributed points?….
6.
What are your costs if you distribute 0 points? ……
7.
By how many percent will your income from the first stage be reduced when you receive from the other group
members a total of:
8.
a.
0 points? …
b.
4 points? ...
c.
15 points? …
At the second stage you distribute the following points to your three other group members: 2, 2, 0. In the third
stage you distribute the following points to your three other group members: 1, 1, 1. What are the total costs of
your distributed points?….
9.
By how many per cent is your second stage income reduced, if you have received the sum of 3 counter-points
from the other group members in stage 3?
25
Appendix 3: Additional Sheet FAV Treatment
Additional information for today’s experiment
This experiment has already been run at two laboratories in Bonn (Germany) and London (UK). The following figure
shows you the contribution behaviour in four selected groups from both locations.
Group 1
Group 2
Group 3
Group 4
max
Average contribution in the group
75%
50%
25%
min
max
75%
50%
25%
min
1
5
10
1
5
10
Period
In these selected graphs you see how much the 4 group members contributed on average to the group project. In the
selected groups 1 and 2 the contributions are high right from the beginning. In the selected groups 3 and 4, average
contribution starts somewhat lower but rises over the course of the experiment.
Some additional numbers from the experiments in Bonn and London:
a)
b)
c)
In the first round of the two previous experiments, 58% to 74% of the contributions were between 10 and 20.
In the last round, up to 53% of the contributions were equal to 20.
Players, whose income was reduced in stage 2, only used counter-points in stage 3 in 29% to 41% of all
possible cases.
All these graphs and numbers are meant to give you some orientation concerning the type of situation you might be
in during the experiment and what you might possibly expect from the other members of your group.
26
Appendix 4: Additional Sheet UNFAV Treatment
Additional information for today’s experiment
This experiment has already been run at two laboratories in Bonn (Germany) and London (UK). The following figure
shows you the contribution behaviour in four selected groups from both locations.
Group 1
Group 2
Group 3
Group 4
max
Average contribution in the group
75%
50%
25%
min
max
75%
50%
25%
min
1
5
10
1
5
10
Period
In these selected graphs you see how much the 4 group members contributed on average to the group project. In the
selected groups 1 and 2 the contributions are rather low and display a decrease over the course of the experiment. In
the selected groups 3 and 4, average contributions are low right from the beginning
Some additional numbers from the experiments in Bonn and London:
a)
b)
c)
In the first round of the two previous experiments, 37% to 55% of the contributions were between 0 and 10.
In the last round, up to 52% of the contributions were equal to 0.
Players, whose income was reduced in stage 2, only used counter-points in stage 3 to reduce the income of
those who had given them points in stage 2. Every time they used counter-points they reduced the income
of those who had given them points in stage 2 by 18% to 25%.
All these graphs and numbers are meant to give you some orientation concerning the type of situation you might be
in during the experiment and what you might possibly expect from the other members of your group.
27
Appendix 5: Belief Elicitation
As noted in the main text, subjects in the belief elicitation had to read the same instructions and answer the same control questions as the participants in the contribution treatments. However, once ztree started subjects were told the following on their
screens:
You have just read the instructions and answered the control questions. This game has already been played in other
laboratories. The participants of those experiments read the same instructions as you and answered the same control
questions. You are not going to play this game today. Instead, your task is to answer the following question:
What do you think, how much did those previous participants contribute on average in the first round to the project?
Your guess………
We have rounded the average down to the next whole number, i.e. the correct answer is a whole number between 0
and 20. If you guess is exactly correct, you will receive 4 Euros. If your guess is wrong, you will receive 0 Euros.
28
Appendix 6: Additional Data Analysis
Table A1: Summary Statistics of Punishment and Counter-punishment across Treatments
BASEFAV
BASEUNFAV
FAV
UNFAV
(1)
Punishment
0.65
0.35
0.92
0.51
(2)
Counter-punishment
0.17
0.11
0.21
0.23
(3)
Counter-punishment (conditional)
0.72
0.78
0.59
1.06
Notes: All figures are means of (counter-)punishment points given per period and subject. (3) is conditional on
subjects being able to allocate counter-punishment. Subjects could only allocate counter-punishment in stage
3 if they had been punished in stage 2.
Parametric Analysis
We have participants interacting over 10 announced periods in fixed groups of four. Moreover
all our data is right-censored since many participants contribute their entire endowment of 20.
Most of our data is also left-censored in that a substantial fraction has kept the entire endowment.
The ideal functional form would therefore be a mixed Tobit model. Unfortunately, such a model
is not available. Random effects Tobit would ignore that observations within groups are not independent. A linear mixed model would ignore that observations are censored. We therefore estimate Tobit models, with lower level 0 and upper level 20, and correct standard errors via bootstrapping, with resampling at the level of entire groups. All models work with 50 repetitions,
which is known to be sufficient for correcting standard errors.
In the comparison between BASEUNFAV and UNFAV, a non-parametric bootstrap (resampling
pairs of dependent and independent variables) yields significant effects. We therefore do not report results from a (semi-)parametric bootstrap resampling residuals (for background see Cameron and Trivedi 2005:360 f.), nor from „wild“ bootstrapping, which randomly assigns error terms
using the procedure developed by (Wu 1986; Mammen 1993), using the parameters proposed by
(Cameron and Trivedi 2005:377); unsurprisingly, showing significance becomes even easier if
one relies on these methods. By contrast, these more advanced methods make a difference in the
comparison between BASEFAV and FAV, which is why we report all three estimates. Since we
only use bootstrapping for correcting standard errors for the dependence of observations, parametric bootstrapping is justified.
29
In the comparison of BASEUNFAV with UNFAV, in all models we find a strong and highly significant negative effect of the belief manipulation. There is no time trend, and no interaction between time and treatment. By contrast, punishment has a highly significant negative main effect:
those who have been punished in the previous period still contribute less than the group average
in the subsequent period. This holds irrespective of the belief manipulation, as shown by the insignificance of the interaction effect.
Table A 2: Comparing BASEUNFAV with UNFAV
model 1
model 2
unfav
-10.708
-10.709
(.001)
(.001)
period
.027
(.896)
unfav*period
model 3
-9.275
(.001)
.163
(.568)
-.263
(.453)
lagged received punishment (lrecpun)
unfav*lrecpun
cons
22.104
(<.001)
21.957
(<.001)
21.218
(<.001)
model 4
-10.972
(.001)
-.421
(.116)
-1.947
(.013)
model 5
-11.583
(.004)
-.474
(.215)
.102
(.796)
-1.949
(.013)
26.022
(<.001)
26.347
(<.001)
model 6
-13.768
(.005)
-.615
(.184)
.304
(.498)
-3.400
(.006)
2.273
(.172)
27.723
(<.001)
depvar: contribution. Tobit, with lower level 0 and upper level 20. Models correct for the dependence of observations per individual and
within groups by bootstrapping, with random draws of entire groups. A non-parametric bootstrap is administered, i.e. pairs of dependent and
their independent variables are randomly selected. p-values in parenthesis. N = 1152. 95 left censored, 494 right censored observations.
In the comparison of BASEFAV with FAV we get a different picture. We have a positive main effect of the belief manipulation in all models. But it is significant only with wild bootstrapping
(models 1, 2, 4), and weakly significant with model 4 (which controls for lagged received punishment). There is, however, a significant interaction between the treatment and the time trend
in model 3. It also is present in wild bootstrapping in models 5 and 6 at conventional significance
levels, and it is weakly significant with alternative bootstrapping methodologies. This indicates
that the belief manipulation stops contributions from quickly eroding. In models 4 and 5, we also
find the negative effect of lagged received punishment, as in the comparison of BASEUNFAV with
UNFAV.
30
Table A 2: Comparing BASEFAVwith FAV
model 1
fav
6.151
(.122)
(.211)
(.004)
period
model 2
6.126
(.123)
(.181)
(.006)
-.584
(.010)
(.001)
(<.001)
model 3
1.071
(.789)
(.745)
(.683)
-1.114
(.001)
(<.001)
(<.001)
.925
(.015)
(.019)
(<.001)
fav*period
lagged received punishment
model 4
7.275
(.085)
(.074)
(.001)
-.939
(<.001)
(<.001)
(<.001)
-1.395
(.038)
(.024)
(<.001)
model 5
3.114
(.502)
(.332)
(.286)
-1.335
(<.001)
(<.001)
(<.001)
.695
(.092)
(.081)
(.014)
-1.369
(.046)
(.021)
(<.001)
13.867
(.002)
(<.001)
(<.001)
16.205
(.001)
(<.001)
(<.001)
fav*lrecpun
cons
7.714
(.023)
(.002)
(<.001)
10.906
(.002)
(<.001)
(<.001)
13.790
(<.001)
(<.001)
(<.001)
model 6
2.612
(.662)
(.433)
(.446)
-1.348
(.001)
(<.001)
(<.001)
.726
(.101)
(.072)
(.018)
-1.610
(.222)
(.101)
(.003)
.405
(.791)
(.709)
(.560)
16.444
(.002)
(<.001)
(<.001)
depvar: contribution. Tobit, with lower level 0 and upper level 20. Models correct for the dependence of observations per individual and within
groups by bootstrapping, with random draws of entire groups. We compare three different bootstrapping techniques, with p-values in parenthesis
and in the following order: non-parametric; parametric over residuals; wild bootstrap. N = 1008. 214 left censored, 277 right censored observations.
31
Preprints 2011
2011/04: Petersen N., The Role of Consent and Uncertainty in the Formation of Customary International Law
forthcoming in: Reexamining Customary International Law, Cambridge, Cambridge University Press, In Press.
2011/03: Petersen N., Antitrust Law and the Promotion of Democracy and Economic Growth
2011/02: Engel C., Besonderes Verwaltungsrecht und ökonomische Theorie
2011/01: Engel C., Zhurakhovska L., Oligopoly as a Socially Embedded Dilemma. An Experiment
Preprints 2010
2010/48: Engel C., An Experimental Contribution to the Revision of the Guidelines on Research and Development Agreements
2010/47: Bade S., Pareto-Optimal Matching Allocation Mechanisms for Boundedly Rational Agents
2010/46: Bade S., Matching Allocation Problems with Endogenous Information Acquisition
2010/45: Bierbrauer F., Boyer P. C., Political competition and Mirrleesian income taxation: A first pass
2010/44: Enders Z., Hakenes H., On the Existence and Prevention of Asset Price Bubbles
2010/43: Weinschenk P., Increasing Workload in a Stochastic Environment
2010/42: Admati A. R., DeMarzo P. M., Hellwig M., Pfleiderer P., Fallacies, Irrelevant Facts, and Myths in the Discussion of
Capital Regulation: Why Bank Equity is Not Expensive
2010/41: Körner T., Schnabel I., Public Ownership of Banks and Economic Growth – The Role of Heterogeneity
2010/40: Ganuza J. J., Jansen J., Too Much Information Sharing? Welfare Effects of Sharing Acquired Cost Information in Oligopoly
2010/39: Weinschenk P., Moral Hazard and Ambiguity
2010/38: Herweg F., Müller D., Weinschenk P., Binary Payment Schemes: Moral Hazard and Loss Aversion
published in: American Economic Review, vol. 100, pp. 2451-2477, 2010.
2010/37: Glöckner A., Engel C., Role Induced Bias in Court: An Experimental Analysis
2010/36: Dang T. V., Hakenes H., Information Disclosure, Intertemporal Risk Sharing, and Asset Prices
2010/35: Traxler C., Burhop C., Poverty and crime in 19th century Germany: A reassessment
2010/34: Dickert S., Herbig B., Glöckner A., Gansen C., Portack R., The More the Better? Effects of Training and Information
Amount in Legal Judgments
2010/33: Towfigh E. V., Glöckner A., Game Over: Empirical Support for Soccer Bets Regulation
2010/32: Glöckner A., Towfigh E. V., Geschicktes Glücksspiel. Die Sportwette als Grenzfall des Glücksspielrechts
published in: JuristenZeitung, vol. 21, pp. 1027-1035, 2010.
2010/31: Hellwig M., Capital Regulation after the Crisis: Business as Usual?
2010/30: Prantl S., The Impact of Firm Entry Regulation on Long-living Entrants
forthcoming in: Small Business Economics, In Press.
2010/29: Engel C., Kurschilgen M., Fairness Ex Ante & Ex Post – The Benefits of Renegotiation in Media Markets
2010/28: Goerg S. J., Walkowitz G., On the Prevalence of Framing Effects Across Subject-Pools in a Two- Person Cooperation
Game
published in: Journal of Economic Psychology, vol. 31, pp. 849-859, 2010.
32
2010/27: Chmura T., Engel C., Englerth M., Pitz T., At the Mercy of the Prisoner Next Door. Using an Experimental Measure of
Selfishness as a Criminological Tool
2010/26: Weinschenk P., Skill Formation under Incomplete Information
2010/25: Brilon S., Job Assignment with Multivariate Skills
2010/24: Lang M., Wambach A., The fog of fraud – mitigating fraud by strategic ambiguity
2010/23: Albanese G., Sorge M. M., The Role of the Judiciary in the Public Decision Making Process
2010/22: Bade S., Electoral Competition with Uncertainty Averse Parties
2010/21: Leifeld P., Haunss S., A Comparison between Political Claims Analysis and Discourse Network Analysis: The Case of
Software Patents in the European Union
2010/20: Burhop C., Gelman S., Transaction costs, liquidity and expected returns at the Berlin Stock Exchange, 1892-1913
2010/19: Hellwig M., Finanzkrise und Reformbedarf [The Financial Crisis and Regulatory Reform]
2010/18: Bierbrauer F., Optimal Income Taxation and Public-Goods Provision with Preference and Productivity Shocks
2010/17: Weinschenk P., Entry and Incumbent Innovation
2010/16: Bierbrauer F., Boyer P. C., The Pareto-Frontier in a simple Mirrleesian model of income taxation
2010/15: Grechenig K., Sekyra M., No Derivative Shareholder Suits in Europe – A Model of Percentage Limits and Collusion
2010/14: Bierbrauer F., On the optimality of optimal income taxation
2010/13: Engel C., An Experimental Contribution to the Theory of Customary (International) Law
2010/12: Leifeld P., Schneider V., Institutional communication revisited: Preferences, opportunity structures and scientic
expertise in policy networks
2010/11: Grechenig K., Nicklisch A., Thöni C., Punishment despite Reasonable Doubt – A Public Goods Experiment with
Uncertainty over Contributions
2010/10: Petersen N., Braucht die Rechtswissenschaft eine empirische Wende?
2010/09: Bade S., Ambiguous Act Equilibria
2010/08: Lehmann S., Volckart O., The Political Economy of Agricultural Protection: Sweden 1887
2010/07: Engel C., Dictator Games: A Meta Study
2010/06: Engel C., Irlenbusch B., Turning the Lab into Jeremy Bentham’s Panopticon. The Effect of Punishment on Offenders
and Non-Offenders
2010/05: Gropp R., Hakenes H., Schnabel I., Competition, Risk-Shifting, and Public Bail-out Policies
2010/04: Slemrod J., Traxler C., Optimal observability in a linear income tax
2010/03: Baumann, Florian; Friehe, Tim; Grechenig, Kristoffel: Switching Consumers and Product Liability: On the Optimality
of Incomplete Strict Liability
2010/02: Bierbrauer F., Hellwig M., Public-Good Provision in a Large Economy
2010/01: Bierbrauer F., Incomplete contracts and excludable public goods
33